Bias Mitigation in ML Classification via Post-Processed Score Functions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Complex machine learning models in financial services struggle to track the importance of individual variables on model behavior, leading to difficulties in mitigating bias, especially in fair lending reviews, as they incorporate a large number of variables and become increasingly difficult to analyze for potential biases impacting protected classes.
Innovation Solution
A method and system that detect bias in machine learning models by comparing distributions using a distance metric, identify contributing input variables through clustering and quantile functions, and mitigate bias through post-processed score functions or fair score approximations, neutralizing or partially neutralizing these variables to reduce bias.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If complex machine learning models are utilized, then predictive accuracy and model performance are improved, but the ability to track variable importance and detect bias deteriorates
Solution Approach 1:
The patent segments the analysis by separating protected attributes from other input variables, allowing independent evaluation of bias impact. This segmentation enables clear identification of which variables contribute to bias while maintaining the full complexity of the ML model for predictive accuracy.
Solution Approach 2:
The patent introduces an intermediary analysis layer that uses surrogate models and mediation techniques to indirectly assess variable importance and bias effects. This intermediary approach simplifies the measurement of bias contribution without requiring direct analysis of the complex original model.
2Reliability
If a large number of variables are incorporated into the model, then the model's predictive capability is improved, but the complexity of analyzing bias contribution increases
Solution Approach 1:
The patent divides the set of input variables into protected attributes and non-protected attributes, enabling selective analysis of bias contribution from specific variable groups while maintaining all variables in the model for predictive capability.
Solution Approach 2:
The patent applies different analysis methods to different variable groups - using mediation analysis for protected attributes and standard importance measures for other variables. This local differentiation simplifies the overall analysis complexity while maintaining comprehensive bias assessment.
3Object-affected harmful factors
If bias mitigation efforts are implemented, then fairness is improved, but the complexity of model processing increases
Solution Approach 1:
The patent performs bias analysis and mitigation planning before final model deployment. By conducting fairness assessments in advance using the full model complexity, the system can simplify processing during actual operation while maintaining fairness commitments.
Solution Approach 2:
The patent modifies model parameters and variable transformations to reduce bias while preserving predictive accuracy. These parameter changes enable the model to achieve fairness goals without requiring fundamental redesign of the processing architecture.
Data Source
AI summary
A computing platform is configured to: (i) train a machine learning model by carrying out a machine learning process on a training data set, wherein the trained machine learning model is configured to (a) receive an input vector comprising respective values for a given set of input variables and (b) based on an evaluation of the received input vector, output a prediction of a given type, (ii) detect bias in the trained machine learning model, (iii) identify one or more input variable groups that contribute to the bias, (iv) mitigate the bias by producing a post-processed version of the trained machine learning model that comprises, for each respective input variable, a respective transformation in place of the respective input variable group, and (v) use the post-processed version of the trained machine learning model to output a given prediction of the given type for a given input vector.